Three open-weight language models shipped inside eight weeks of each other this summer, and none of them came from the usual suspects. Tencent’s Hunyuan team released Hy3 on July 6, 2026. Moonshot AI followed with Kimi K3’s API on July 16 and its full weights on July 27. Then Zhipu AI’s Z.ai division dropped GLM-5.3-Flash on August 26, pairing a roughly 1M-token context window with the cheapest per-token pricing of the three. Anyone shopping for an open source LLM in September 2026 now has to pick between a 295-billion-parameter reasoning model, a natively multimodal 320-billion-parameter workhorse, and a 2.8-trillion-parameter model that Moonshot calls the first open 3-trillion-class release. The gap between them is not small: on published list pricing, Kimi K3 costs roughly 20 times more per input token than GLM-5.3-Flash through common inference marketplaces. This piece breaks down the specs, the benchmarks, the licensing fine print, and which model actually fits which job.
What Are Tencent Hy3, GLM-5.3-Flash, and Kimi K3?
All three are Mixture-of-Experts (MoE) models, meaning each only activates a fraction of its total parameters for any given token, which keeps inference cheaper than a dense model of the same size would be. Beyond that shared architecture choice, the three labs took different bets on what an open-weight model should optimize for in late 2026.
Tencent Hy3 is the third generation of the Hunyuan family, built by what the company now calls the Hy Team, and it was shipped specifically as a reasoning-and-coding model with an unusually clean license: Apache 2.0, with no field-of-use or geographic restrictions, according to Tencent’s own release documentation. Hy3 is a 295-billion-parameter MoE with 192 experts, 8 of which activate per token, for roughly 21 billion active parameters at inference time.
GLM-5.3-Flash comes from Z.ai (Zhipu AI), and it is pitched as a natively multimodal model rather than a pure text reasoner. It carries 320 billion total parameters with about 18 billion active per token, an MIT license, and a context window Z.ai markets as 1M tokens (providers report the exact figure anywhere from 1,048,576 to 1,310,720 tokens depending on how they round it, per Cloudflare’s Workers AI model page).
Kimi K3 is Moonshot AI’s biggest swing: a 2.8-trillion-parameter MoE with 896 experts, 16 of which activate per token, for around 104 billion active parameters. Moonshot’s own marketing calls it the world’s first open 3-trillion-class model, and its Hugging Face model card lists benchmark numbers that beat both Hy3 and GLM-5.3-Flash on several shared tests, at the cost of a considerably more complicated license.
The timing of all three releases is not a coincidence. Chinese AI labs have spent 2026 racing to ship open-weight models that close the gap with closed frontier systems from OpenAI, Anthropic, and Google, and DeepSeek’s V4 family set the pace earlier in the year by proving an open source LLM could compete on price and reasoning benchmarks at the same time. Tencent, Z.ai, and Moonshot each responded with a different specialization rather than a straight DeepSeek clone: Tencent leaned into licensing simplicity, Z.ai leaned into multimodality and cost, and Moonshot leaned into raw scale. That divergence is exactly why a single “best open source LLM” answer doesn’t hold up once you look past the top-line benchmark numbers — the three models are solving for different constraints, not competing head-to-head on identical terms.
Specs at a Glance: Parameters, Context, and Licensing
Here is the full spec sheet for all three open source LLMs as of September 2, 2026, pulled from each lab’s release documentation and model cards.
| Spec | Tencent Hy3 | GLM-5.3-Flash | Kimi K3 |
|---|---|---|---|
| Developer | Tencent (Hy Team / Hunyuan) | Z.ai (Zhipu AI) | Moonshot AI |
| Release date | July 6, 2026 | August 26, 2026 | July 16, 2026 (API); July 27, 2026 (weights) |
| Architecture | Mixture-of-Experts | Mixture-of-Experts, natively multimodal | Mixture-of-Experts |
| Total parameters | 295B | 320B | 2.8T |
| Active parameters/token | ~21B (8 of 192 experts) | ~18B | ~104B (16 of 896 experts) |
| Context window | 256K tokens | ~1M tokens (1,048,576–1,310,720 by provider) | 1,000,000 tokens |
| License | Apache 2.0, no restrictions | MIT | Custom “Kimi K3 License” (modified MIT base) |
| Weights size | 598GB (BF16) / 300GB (FP8) | Not disclosed in BF16 or FP8 totals by Z.ai | ~1.4TB (MXFP4 quantized) |
| Multimodal | Text/reasoning focused | Yes, natively multimodal | Native multimodal claimed at launch |
| Hosting | Hugging Face, ModelScope, GitCode, CNB | Hugging Face, Cloudflare Workers AI, 7+ inference providers | Hugging Face (moonshotai/Kimi-K3) |
| Commercial use threshold | None | None | Contract required above $20M annual revenue |
The parameter gap is the headline number here: Kimi K3’s 2.8 trillion total parameters are roughly 9.5 times larger than Hy3’s 295 billion and 8.75 times larger than GLM-5.3-Flash’s 320 billion. That scale shows up in benchmark scores, but it also shows up in the weights download, which at roughly 1.4TB even after MXFP4 quantization is not something you fit on a couple of consumer GPUs.
Two other rows in that table deserve a second look before you commit to a model. The release-date spread matters because Kimi K3’s July timeline means it has had roughly seven weeks longer than GLM-5.3-Flash to accumulate real-world usage reports and community fine-tunes, which is worth factoring in if you’re weighing maturity alongside raw specs. And the commercial-use threshold row is the one line item that can quietly change your total cost of ownership: a team that ignores Kimi K3’s $20 million revenue trigger during a pilot project can find itself needing a licensing conversation right as the project is ready to scale into production.
Architecture Deep Dive: How Each Mixture-of-Experts Model Works
Hy3’s 192-Expert Design
Hy3 routes each token through 8 of its 192 experts, and Tencent also built in a separate 3.8-billion-parameter multi-token prediction layer, according to model documentation aggregated by AI/TLDR’s model tracker. That layer lets the model predict several tokens ahead in a single pass, which is part of why Tencent positions Hy3 as competitive on long-horizon reasoning despite its comparatively modest 256K context window. The tradeoff is that Hy3’s context ceiling is roughly a quarter of what GLM-5.3-Flash and Kimi K3 offer, which matters for anyone processing full codebases or long legal documents in a single pass.
Kimi K3’s 2.8 Trillion Parameters
Kimi K3 activates 16 of 896 experts per token, landing at approximately 104 billion active parameters, per Moonshot’s own model card. That is roughly five times the active-parameter count of Hy3 or GLM-5.3-Flash, even though all three are MoE designs. The tradeoff shows up immediately in serving cost: more active parameters per token means more compute per generated token, which is a big part of why Kimi K3 is priced well above the other two on third-party inference marketplaces.
GLM-5.3-Flash’s Multimodal MoE Approach
GLM-5.3-Flash sits between the other two in both total and active parameters, at 320 billion total and roughly 18 billion active per token, but its architecture pitch is different from either Hy3 or Kimi K3: Z.ai built it as a natively multimodal MoE from the ground up rather than bolting vision or audio input onto a text-first model after the fact. That native multimodal design, combined with a roughly 1M-token context window, is what lets GLM-5.3-Flash sit on Cloudflare’s Workers AI platform as a general-purpose model rather than a specialist reasoning engine, and it’s part of why Z.ai can undercut both Hy3 and Kimi K3 on price — a smaller active-parameter count per token means the model is cheaper to serve at scale even before accounting for any promotional discount.
Benchmark Showdown: GPQA, Terminal-Bench, and Coding Scores
Benchmark reporting across open-weight model releases is inconsistent, and no single source has run all three models through an identical evaluation harness. The table below uses the scores each lab or an independent tracker has published, and it flags gaps honestly rather than filling them with estimates.
| Benchmark | Tencent Hy3 | GLM-5.3-Flash | Kimi K3 |
|---|---|---|---|
| GPQA Diamond | 90.4% | Not disclosed | 93.5% (Hugging Face); 92.9% per vals.ai |
| SWE-bench Verified | 78% | Not disclosed | Not disclosed |
| SWE-bench Pro | 57.9% | Not disclosed | Not disclosed |
| Terminal-Bench 2.1 | Not disclosed (reports LHTB instead) | 84.3 | 88.3 |
| HLE, no tools | 53.2% | 39.9 | 43.5–46.9 (source-dependent) |
| HLE, with tools | Not disclosed | Not disclosed | 56.0 |
| Long-Horizon Terminal Bench | 28.8% | Not disclosed | Not disclosed |
| BrowseComp | 84.2 | Not disclosed | Not disclosed |
Where the three models do overlap, Kimi K3 comes out ahead: 93.5% on GPQA Diamond versus Hy3’s 90.4%, and 88.3 on Terminal-Bench 2.1 versus GLM-5.3-Flash’s 84.3, a 4-point gap on that agentic coding test. According to a breakdown from MarkTechPost’s coverage of the GLM-5.3-Flash launch, Z.ai’s own comparison table put GLM-5.3-Flash’s Terminal-Bench 2.1 score close behind Anthropic’s Claude Opus 4.8 (85.0) and OpenAI’s GPT-5.6 Terra (87.4), which suggests the open-weight field as a whole has closed most of the gap with closed frontier models on that particular test. Hy3 does not report a Terminal-Bench 2.1 number at all, instead publishing its own Long-Horizon Terminal Bench score of 28.8%, a different and apparently harder test that makes a direct comparison impossible.
It’s worth being upfront about how much of this table is self-reported. Hy3’s numbers come from Tencent’s own release documentation and independent aggregators like AI/TLDR and Digital Applied; GLM-5.3-Flash’s come from Z.ai’s launch announcement as covered by MarkTechPost, plus independent tracker The Model Gap; and Kimi K3’s come from Moonshot’s Hugging Face model card, cross-checked against the vals.ai-sourced numbers on The Model Gap. That’s three independent, named sources feeding into this table, but none of them ran a single unified benchmark suite across all three models at once. Until an independent third party does exactly that, this table should be read as “what each lab chose to publish,” not a fully neutral bake-off.
Licensing Compared: Apache 2.0 vs MIT vs the Kimi K3 License
Licensing is where the three models diverge the most, and it is the detail most likely to bite a team that skips straight to the benchmark table.
- Tencent Hy3: Apache 2.0, described in Tencent’s release notes as having no field-of-use or geographic restrictions. This is about as permissive as an open source LLM license gets — fine-tune it, redistribute it, sell a product built on it, no contract required.
- GLM-5.3-Flash: MIT license, also unrestricted for commercial use, per Z.ai’s own announcement and confirmed independently by Cloudflare’s Workers AI model listing.
- Kimi K3: A custom document Hugging Face lists simply as “kimi-k3,” not standard MIT despite early coverage calling it “MIT-based.” Per Wikipedia’s entry on Kimi, any company with annual revenue above $20 million must negotiate a contract with Moonshot AI before offering Kimi K3 as a hosted service, and high-scale products must attribute the model.
That $20 million revenue threshold is a real planning constraint, not boilerplate. A startup under that ceiling can use Kimi K3’s weights freely, including commercially, but the moment the company crosses it, Kimi K3 stops being a drop-in dependency and becomes a contract negotiation. Hy3 and GLM-5.3-Flash carry no equivalent tripwire, which is a meaningful reason some engineering teams default to one of those two even when Kimi K3 scores higher on paper.
This is also a broader industry shift worth flagging. DeepSeek built its reputation partly on shipping plain MIT-licensed weights with no revenue gates at all, and for a while that set the expectation for what an open source LLM release should look like. Moonshot’s revenue-gated approach with Kimi K3 is a departure from that norm, and it puts K3 in a slightly different category from a strict licensing standpoint — closer to a “free for most, commercial for scale” model than a fully unrestricted open release. Whether that becomes the new normal for trillion-parameter open weights, or whether Hy3 and GLM-5.3-Flash’s unrestricted approach wins out, is one of the more interesting open questions in the open-weight AI market heading into the rest of 2026.
Pricing and Inference Costs: The 20x Gap Explained
None of these three labs publishes a single universal API price — pricing depends on whether you use the lab’s own hosted endpoint, a self-hosted deployment, or a third-party inference marketplace. Here is what is publicly documented as of this writing.
| Pricing detail | Tencent Hy3 | GLM-5.3-Flash | Kimi K3 |
|---|---|---|---|
| Official API, input/1M tokens | Not publicly listed by Tencent | $0.15 | Not directly published by Moonshot |
| Official API, output/1M tokens | Not publicly listed by Tencent | $0.50 | Not directly published by Moonshot |
| Cached input/1M tokens | N/A | $0.03 | N/A |
| Launch promo pricing | N/A | $0.075 in / $0.25 out, through Sept 9, 2026 | N/A |
| Third-party marketplace, input/1M | Varies by host | ~$0.075–$0.15 (OpenRouter and others) | $3.00 (CometAPI listing) |
| Third-party marketplace, output/1M | Varies by host | ~$0.25–$0.50 | $15.00 (CometAPI listing) |
| Primary cost path | Self-host (open weights) | Hosted API or self-host | Hosted API (third-party) or self-host |
The 20x gap referenced in this article’s headline compares GLM-5.3-Flash’s official $0.15-per-million-token input price against Kimi K3’s $3.00-per-million-token input price as listed on the CometAPI inference marketplace: $3.00 divided by $0.15 is exactly 20. The same math holds on output tokens too, at $15.00 versus $0.75 depending on which GLM-5.3-Flash tier you compare against (the ratio narrows to roughly 8x if you compare against Z.ai’s $0.50 output-token figure directly). Either way, GLM-5.3-Flash is the clear budget option of the three, and Kimi K3’s larger active-parameter count carries a real cost premium wherever it is hosted. Tencent has not published an equivalent hosted-API price sheet for Hy3, which pushes most adopters toward self-hosting it instead, an option this piece covers further down.
To put those per-token prices into a workload context: a pipeline processing 10 million input tokens a day would cost roughly $1.50 a day on GLM-5.3-Flash’s official rate, versus roughly $30 a day for the same volume through Kimi K3 on CometAPI’s listed price — a difference that compounds fast once you’re running that pipeline continuously across a month or a quarter. That gap doesn’t mean GLM-5.3-Flash is automatically the right call for every workload; it means the cost of testing Kimi K3’s higher benchmark scores against your own eval set should be weighed against what that 20x price premium actually buys you in output quality for your specific task.
Context Windows and Long-Document Performance
Context window is one of the few specs all three labs report cleanly, and it is also one of the more consequential differences for real workloads. Hy3 tops out at 256K tokens, which is enough for a large but not enormous codebase or a few hundred pages of text. GLM-5.3-Flash and Kimi K3 both advertise roughly 1M tokens, though providers list GLM-5.3-Flash’s exact figure inconsistently — Cloudflare’s own model page states 1,048,576 tokens, while another aggregator lists 1,310,720 tokens for the same model, a discrepancy likely tied to how each provider counts special tokens or output allowances.
For teams building retrieval-heavy applications, that roughly 4x context gap between Hy3 and the other two models is often more decisive than a few benchmark points. A million-token window can hold an entire mid-size monorepo or a full legal contract set in one pass; 256K generally cannot. GLM-5.3-Flash’s output cap is documented at 128K tokens even with its 1M-token input window, which is worth checking against your use case if you need very long generated output rather than just long input context.
A practical example: a legal team running discovery across a 400-page contract bundle plus supporting correspondence could plausibly land north of 300,000 tokens once everything is tokenized, which still fits comfortably inside GLM-5.3-Flash’s or Kimi K3’s window but would force Hy3 to chunk the document into two or more passes. Chunking isn’t fatal — plenty of production RAG pipelines do it by necessity — but it adds an extra layer of orchestration and retrieval logic that a single-pass model doesn’t need.
Self-Hosting Requirements: VRAM, Quantization, and Hardware
All three models ship as open weights, which means self-hosting is on the table for teams with the GPU budget. Hy3’s documented footprint is 598GB in BF16 or 300GB after FP8 quantization, according to spec sheets published alongside its release — both figures point toward a multi-GPU server-class deployment rather than a single workstation card. Kimi K3’s weights land around 1.4TB even in MXFP4 quantization, the lowest-precision format Moonshot shipped, which puts it firmly in data-center territory; running it well requires a cluster of high-memory accelerators, not a single node. Z.ai has not published an equivalent total-gigabyte figure for GLM-5.3-Flash’s BF16 or FP8 builds, though its comparatively smaller 320B parameter count suggests a footprint closer to Hy3’s than to Kimi K3’s.
A representative vLLM launch command for a quantized Hy3 deployment looks like this (the same pattern applies to calling GLM-5.3-Flash through its API instead of self-hosting, if your GPU budget doesn’t stretch to a multi-node cluster):
vllm serve tencent/Hy3-FP8
--tensor-parallel-size 8
--max-model-len 262144
--quantization fp8
--gpu-memory-utilization 0.92
--served-model-name hy3None of the three models are realistic candidates for a single consumer GPU, even quantized. If your hardware budget tops out at one or two 24GB or 48GB cards, all three vendors’ hosted APIs (or a third-party inference marketplace) are the only practical path, and GLM-5.3-Flash’s pricing makes it the cheapest of the three to run that way at any real scale.
It’s also worth planning for the fact that self-hosting a Mixture-of-Experts model isn’t the same operational job as self-hosting a dense model of comparable total size. Expert routing adds its own scheduling and memory-bandwidth considerations, and serving frameworks like vLLM, SGLang, and TensorRT-LLM handle MoE routing differently, with performance characteristics that can vary meaningfully between them. Teams without existing MoE-serving experience should budget time for tuning tensor-parallel and expert-parallel settings before assuming a self-hosted deployment will hit the throughput numbers a lab’s own benchmarks imply.
Real-World Use Cases and Early Adopters
Concrete deployment patterns are already forming around each model’s specific tradeoffs:
- Regulated on-prem deployments. Hy3’s Apache 2.0 license with no geographic or field-of-use restriction makes it a straightforward pick for banks, healthcare providers, or government contractors that need to run inference entirely inside their own infrastructure without a licensing review.
- Edge and serverless inference. GLM-5.3-Flash’s presence on Cloudflare’s Workers AI platform means teams already running on Cloudflare’s edge network can call it without standing up their own GPU fleet at all.
- Long-context document review. Legal, compliance, and due-diligence teams processing full contract sets or discovery documents lean toward GLM-5.3-Flash or Kimi K3 for their roughly 1M-token windows rather than Hy3’s 256K ceiling.
- Frontier-level coding agents. Kimi K3’s 88.3 on Terminal-Bench 2.1 and its status as Moonshot’s flagship make it the natural pick for teams building autonomous coding agents who need the highest raw capability an open-weight model currently offers, and who fall under the $20 million revenue threshold or are prepared to negotiate a contract if they don’t.
- Cost-sensitive high-volume APIs. Customer support bots, summarization pipelines, and other high-throughput but lower-complexity workloads are well suited to GLM-5.3-Flash’s $0.15/$0.50 official pricing, which undercuts Kimi K3 by a wide margin for tasks that do not need frontier-level reasoning.
These five patterns aren’t mutually exclusive, either. A single organization might reasonably run GLM-5.3-Flash for its customer-facing support bot, Hy3 for an internal on-prem document search tool subject to compliance review, and pilot Kimi K3 for a coding-agent prototype before deciding whether the licensing conversation with Moonshot is worth having at production scale. Treating this as a portfolio decision rather than a single model choice is increasingly how teams with more than one AI workload are approaching the open-weight market in late 2026.
How Each Model Stacks Up Against DeepSeek V4 and Qwen3.8
DeepSeek V4 and Qwen3.8 remain the two most widely deployed open-weight reference points in the market, and all three of this article’s models are implicitly being benchmarked against them by developers even where labs have not published a direct head-to-head table. Hy3, GLM-5.3-Flash, and Kimi K3 all shipped after DeepSeek V4 Pro’s well-documented August 2026 release, and each is explicitly positioned as a challenger to it on price, context, or raw capability rather than as a wholesale replacement for a single dominant model. Kimi K3’s GPQA Diamond score of 93.5% and its 2.8-trillion-parameter scale put it in a different weight class than most Qwen3.8-series models, which typically run in the hundreds of billions of parameters rather than trillions. GLM-5.3-Flash, by contrast, competes more directly with DeepSeek’s Flash-tier releases on price and context window rather than trying to out-benchmark the largest reasoning-focused models.
None of the sources reviewed for this piece published a single table placing all five models — Hy3, GLM-5.3-Flash, Kimi K3, DeepSeek V4, and Qwen3.8 — side by side on identical benchmarks, so treat any claim of a definitive ranking across all five with skepticism until an independent, uniform evaluation exists.
What is clear is that the open-weight tier of the AI market has gotten crowded fast. A year ago, choosing an open-weight model meant picking between a handful of Llama and Mistral variants and one or two DeepSeek releases. Now a team evaluating open weights in September 2026 has to weigh Hy3, GLM-5.3-Flash, and Kimi K3 against DeepSeek V4, DeepSeek V4-Flash, Qwen3.5 and Gemma 4-class models, GLM-5.2 (already benchmarked against an earlier Kimi K3 pairing in our Kimi K3 vs Qwen3.8-Max vs GLM-5.2 comparison), MiniMax M3, and Nvidia’s Nemotron line, often with overlapping release windows and inconsistent benchmark reporting across all of them. That’s good news for buyers in terms of price competition, but it makes a rigorous, apples-to-apples bake-off on your own workload more valuable than trusting any single vendor’s leaderboard position.
Migration Guide: Moving From a Closed API to an Open-Weight Model
Switching a production workload from a closed API (or from one of these three models to another) follows a similar checklist regardless of which one you land on. The mechanics of swapping an endpoint are simple; the risk is almost always in output-quality drift and unaccounted-for licensing or cost surprises that only surface after you’re already in production.
Here’s the sequence that keeps that risk manageable:
- Audit your current prompt and output patterns, including average token counts, so you know whether Hy3’s 256K context is enough or whether you need GLM-5.3-Flash’s or Kimi K3’s ~1M window.
- Build a small evaluation set (50-200 real prompts) from your actual production traffic rather than relying solely on published GPQA or Terminal-Bench scores, since none of the three publish identical benchmark suites.
- Check your organization’s trailing annual revenue against Kimi K3’s $20 million threshold before committing engineering time to it; if you’re near or above that line, start the Moonshot licensing conversation early or default to Hy3 or GLM-5.3-Flash instead.
- Decide between self-hosting and a hosted API based on your GPU budget: sub-500GB VRAM budgets rule out self-hosting Kimi K3 outright, and even Hy3’s 300GB FP8 build needs a multi-GPU node.
- If self-hosting, quantize to FP8 or MXFP4 first and benchmark latency under your real concurrency, not a single-request test.
- Run the new model in shadow mode against your existing API for at least a week before cutting over, comparing both output quality and cost per request.
- Keep a rollback path to your prior model or API active until the new deployment has cleared at least one full billing cycle without incident.
Teams moving away from a closed API for the first time should also budget extra time for prompt re-tuning. System prompts, few-shot examples, and tool-calling formats optimized for a closed model like GPT-5.6 or Claude Opus 5 rarely transfer cleanly to an open-weight model without adjustment, and that re-tuning work is usually a bigger time sink than the infrastructure migration itself.
Pros and Cons of Each Model
| Model | Pros | Cons |
|---|---|---|
| Tencent Hy3 | Cleanest license (Apache 2.0, zero restrictions); strong SWE-bench Verified (78%); FP8 build shrinks to 300GB | Smallest context window (256K); no published Terminal-Bench 2.1 score; no official hosted API pricing |
| GLM-5.3-Flash | Cheapest official pricing ($0.15/$0.50 per 1M); ~1M context; natively multimodal; live on Cloudflare Workers AI | No published GPQA or SWE-bench scores; lower HLE score (39.9) than the other two; weights footprint not disclosed |
| Kimi K3 | Highest GPQA Diamond (93.5%) and Terminal-Bench 2.1 (88.3) of the three; largest scale (2.8T params); 1M context | ~20x pricier per token on third-party marketplaces; $20M revenue license trigger; ~1.4TB weights need data-center hardware |
Adoption signals this early are thin for all three, which is itself worth noting. None of the sources reviewed for this piece published Hugging Face download counts, GitHub star totals, or named enterprise customers for Hy3 or Kimi K3, and the strongest concrete adoption data point available is GLM-5.3-Flash’s listing across seven inference providers plus its live availability on Cloudflare’s Workers AI platform. That’s a meaningfully different signal than a download counter, but it does confirm at least one of the three has already cleared the bar of “multiple independent platforms trust this enough to host it,” which isn’t nothing for a model that shipped less than two weeks before this article was published.
Which Open Source LLM Should You Choose?
The right pick depends heavily on which constraint binds hardest for your team: license risk, context length, raw benchmark ceiling, or per-token cost. There isn’t a universal answer, which is exactly why the “best open source LLM” framing that dominates most roundup articles tends to break down once you look at a specific deployment rather than a leaderboard.
- Pick Tencent Hy3 if you need Apache 2.0’s zero-restriction license for a regulated or on-prem deployment and your context needs fit inside 256K tokens.
- Pick GLM-5.3-Flash if cost per token is your primary constraint, you need multimodal input, or you’re already running on Cloudflare’s edge infrastructure.
- Pick Kimi K3 if you need the highest available open-weight benchmark scores for an agentic coding or reasoning product and either your revenue is under $20 million or you’re prepared to negotiate Moonshot’s commercial terms.
- Pick GLM-5.3-Flash over Kimi K3 if you’re a startup approaching the $20 million revenue line and want to avoid a licensing renegotiation down the road.
- Pick Hy3 over GLM-5.3-Flash if SWE-bench-style coding correctness matters more to your product than raw context length, since Hy3’s 78% SWE-bench Verified score is the only one of the three that’s publicly documented on that specific test.
The Verdict: Data-Backed Recommendation
None of these three open source LLMs is a strict upgrade over the other two — each is optimized for a different constraint, and the benchmark gaps that do exist are not uniform across every test. Kimi K3 wins the raw capability argument on every benchmark where a direct comparison is possible (GPQA Diamond, Terminal-Bench 2.1), but it comes with a licensing tripwire and a per-token cost that is roughly 20 times GLM-5.3-Flash’s official rate on third-party marketplaces. GLM-5.3-Flash wins on price, context window, and license simplicity, at the cost of a lower HLE score and the absence of any published GPQA or SWE-bench number, which makes its true reasoning ceiling harder to verify. Hy3 sits in between: the most permissive license of the three and a genuinely strong SWE-bench Verified score, undercut by a context window a quarter the size of its two rivals.
For most engineering teams evaluating an open source LLM today, GLM-5.3-Flash is the pragmatic default given its combination of MIT licensing, documented pricing, and million-token context. Teams building a coding or reasoning product where benchmark ceiling matters more than budget should test Kimi K3 directly against their own eval set before committing, keeping the $20 million revenue threshold in mind. And any team with a hard compliance or on-prem requirement should start with Hy3, since Apache 2.0 removes an entire category of legal review that the other two licenses don’t.
The single number worth remembering from this entire comparison is the 20x figure: GLM-5.3-Flash’s $0.15-per-million-token input price against Kimi K3’s $3.00 on third-party marketplaces. That gap alone should push most cost-conscious teams to start their evaluation with GLM-5.3-Flash, reserve Kimi K3 for workloads where its extra 3-4 benchmark points on GPQA and Terminal-Bench 2.1 demonstrably move a business metric, and keep Hy3 on the shortlist specifically for its licensing simplicity rather than its raw scores. None of the three is wrong to pick — they’re just answering different questions, and the data above should make it clear which question matches your own constraints.
Frequently Asked Questions
Is Tencent Hy3 free to use commercially?
Yes. Hy3 ships under Apache 2.0 with no field-of-use or geographic restrictions, so commercial use, fine-tuning, and redistribution are all permitted without a separate contract.
What license does Kimi K3 use, and is it really MIT?
Hugging Face lists it under a custom document called “kimi-k3,” not a standard MIT license, even though early coverage described it as MIT-based. Companies with more than $20 million in annual revenue must negotiate a contract with Moonshot AI before offering Kimi K3 as a hosted service.
Which of the three has the largest context window?
GLM-5.3-Flash and Kimi K3 both advertise roughly 1M tokens, though exact figures vary slightly by provider. Hy3 tops out at 256K tokens, about a quarter of the other two.
Why is Kimi K3 so much more expensive than GLM-5.3-Flash?
Kimi K3 activates roughly 104 billion parameters per token versus GLM-5.3-Flash’s 18 billion, which drives up compute cost per generated token. On the CometAPI inference marketplace, that translates to $3.00 per million input tokens for Kimi K3 versus $0.15 for GLM-5.3-Flash’s official Z.ai rate, a 20x gap.
Can I self-host any of these three models on a single GPU?
No. Even quantized, Hy3’s FP8 build is roughly 300GB and Kimi K3’s MXFP4 build is roughly 1.4TB, both of which require multi-GPU, data-center-class hardware. GLM-5.3-Flash’s exact weight size hasn’t been published by Z.ai, but its 320B parameter count suggests a similar multi-GPU requirement.
Which model scores highest on coding benchmarks?
Kimi K3 reports the highest Terminal-Bench 2.1 score (88.3) among the two models that publish one, ahead of GLM-5.3-Flash’s 84.3. Hy3 does not publish a Terminal-Bench 2.1 score, but reports a 78% SWE-bench Verified score, a different test that the other two do not publicly report.
Is GLM-5.3-Flash actually multimodal, or just text?
Z.ai describes GLM-5.3-Flash as natively multimodal, and it is listed as such on its Hugging Face model card and on hosting platforms including Cloudflare Workers AI.
Do any of these three models beat DeepSeek V4 or Qwen3.8 outright?
No single published source puts all five models side by side on identical benchmarks, so a definitive ranking across all five cannot be verified. Kimi K3’s 2.8-trillion-parameter scale and 93.5% GPQA Diamond score put it in a different class than most Qwen3.8 variants, but that is a scale comparison, not a confirmed head-to-head benchmark win.
